基于深度学习实例分割的室内环境物体语义建图

Object-oriented semantic mapping for indoor environments based on instance segmentation with deep learning

为提升家庭服务机器人面向物体语义建图的精确性,提出一种改进的语义建图方法。该方法基于深度学习对单帧图像进行实例分割,利用其结果指导基于几何关系的聚类,使聚类结果既能保留深度学习方法识别的物体语义信息,又能贴近物体的几何边界。针对现有方法所建地图中物体标签缺乏可靠性的问题,提出一种语义地图优化方法,对地图中距离相近、重合度较高、类别相似的重复物体进行合并,并删除错误物体。实验验证表明,提出的方法能够提升物体边界分割的准确性,得到的物体数目更加准确。

To enhance the fidelity of the object-oriented semantic maps of domestic service robots,this paper proposed an improved semantic map building method.The method performed instance segmentation on the images with a deep learning method and instructed a geometric relationship-based clustering method to segment objects from the point clouds.The clustering results not only preserved the semantic information identified by the deep learning method,but also described geometric boundaries of objects more precisely.To solve the problem of the lack of reliability for object labels in existing approaches,this paper also proposed a semantic map optimization process.It merged instances which had close distances,high overlapping rates,and similar label scores.Meanwhile,it removed false objects.The experimental results show that the proposed method can improve the quality of the map built by the system in terms of the accuracy of the boundaries and the number of instances.